Why this record is void
Weight-allocation bug in sweep/cross_sectional.py: assets outside the point-in-time universe scored -inf and therefore sorted first, so the entire short leg of every long_short variant (45 of 90) was allocated to non-investable assets whose returns are zeroed. Verified by replay: the stored returns matrix reproduces exactly under the buggy allocator (max abs diff 2.6e-08) and diverges materially under the fixed one (correlation 0.54-0.69; one variant moves from 0.187 to 0.468 annualised Sharpe). The mean_touch (long-only) branch was unaffected, but the corrupted variants are counted in n_trials, in the expected-maximum-Sharpe null and in the PBO combinatorics, so the primary metric is not valid. Pre-registration and its Bitcoin timestamp remain valid; only the execution failed. Fixed and covered by tests/test_cross_sectional_weights.py; re-registered under new IDs.
What was sealed, and the verdict
Pre-registered test: does the best of 90 cross-sectional momentum variants clear DSR 0.95? Verdict: NO. DSR 0.809 — the closest any sweep in this registry has come, and still short.
Noticed afterwards: a hypothesis, not a finding
96% of variants were positive with a median annualised Sharpe of 0.51 and a best of 0.98 against an expected-under-null of 0.64. The family behaves as one thing rather than as a spread of better and worse rules.
PBO is 0.479, near the 0.5 coin-flip line. Choosing among these 90 variants carries almost no information out of sample. That is consistent with a uniform family-wide effect and inconsistent with some variants genuinely being better.
The effect comes from the stable end of the universe, not the volatile end. Splitting by symbol lifetime: 81 long-lived symbols (>40,000 bars) give a net Sharpe of 0.72, while 55 short-lived symbols (<20,000 bars) give 0.04. A spurious altcoin-volatility artefact would show the opposite, and the effect is stronger where liquidity is better.
Post hoc observation, carried forward to the replacement sweeps: the mean_touch exit rule is long-only with gross exposure 2.0, not dollar-neutral, so it is not market-neutral and does inherit the crypto trend. In all three sweeps the highest-Sharpe variant came from this branch. In SW-0034 that variant compounds to 0.22x of initial capital with a 100% maximum drawdown despite an annualised Sharpe of 0.989, losing between 37% and 53% in each of four hours during the LUNA/UST collapse (2022-05-11 to 2022-05-13; verified against BTCUSDT bar timestamps, not a data error). Sharpe and the deflated Sharpe ratio are computed from arithmetic moments and do not see the capital path.
Notes
This is the only family in 33 records to show a coherent signal, and it is also the only one with substantial published academic support in crypto. The test cannot confirm it: the detection limit is around 1.6 annualised Sharpe and the observed effect is roughly 0.7.
SW-0035, embargoed until 2029-06-30, is the holdout test — and its pre-registration states up front that it is underpowered. Confirming a true Sharpe of 0.72 at z=2 with 80% power needs roughly 136,000 hourly bars, about 15.6 years. Four years gives about 29%.